1 /* 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * The Mach Operating System project at Carnegie-Mellon University. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by the University of 19 * California, Berkeley and its contributors. 20 * 4. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 37 * $FreeBSD$ 38 */ 39 40 /* 41 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 42 * All rights reserved. 43 * 44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 45 * 46 * Permission to use, copy, modify and distribute this software and 47 * its documentation is hereby granted, provided that both the copyright 48 * notice and this permission notice appear in all copies of the 49 * software, derivative works or modified versions, and any portions 50 * thereof, and that both notices appear in supporting documentation. 51 * 52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 55 * 56 * Carnegie Mellon requests users of this software to return to 57 * 58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 59 * School of Computer Science 60 * Carnegie Mellon University 61 * Pittsburgh PA 15213-3890 62 * 63 * any improvements or extensions that they make and grant Carnegie the 64 * rights to redistribute these changes. 65 */ 66 67 /* 68 * GENERAL RULES ON VM_PAGE MANIPULATION 69 * 70 * - a pageq mutex is required when adding or removing a page from a 71 * page queue (vm_page_queue[]), regardless of other mutexes or the 72 * busy state of a page. 73 * 74 * - a hash chain mutex is required when associating or disassociating 75 * a page from the VM PAGE CACHE hash table (vm_page_buckets), 76 * regardless of other mutexes or the busy state of a page. 77 * 78 * - either a hash chain mutex OR a busied page is required in order 79 * to modify the page flags. A hash chain mutex must be obtained in 80 * order to busy a page. A page's flags cannot be modified by a 81 * hash chain mutex if the page is marked busy. 82 * 83 * - The object memq mutex is held when inserting or removing 84 * pages from an object (vm_page_insert() or vm_page_remove()). This 85 * is different from the object's main mutex. 86 * 87 * Generally speaking, you have to be aware of side effects when running 88 * vm_page ops. A vm_page_lookup() will return with the hash chain 89 * locked, whether it was able to lookup the page or not. vm_page_free(), 90 * vm_page_cache(), vm_page_activate(), and a number of other routines 91 * will release the hash chain mutex for you. Intermediate manipulation 92 * routines such as vm_page_flag_set() expect the hash chain to be held 93 * on entry and the hash chain will remain held on return. 94 * 95 * pageq scanning can only occur with the pageq in question locked. 96 * We have a known bottleneck with the active queue, but the cache 97 * and free queues are actually arrays already. 98 */ 99 100 /* 101 * Resident memory management module. 102 */ 103 104 #include <sys/param.h> 105 #include <sys/systm.h> 106 #include <sys/lock.h> 107 #include <sys/malloc.h> 108 #include <sys/mutex.h> 109 #include <sys/proc.h> 110 #include <sys/vmmeter.h> 111 #include <sys/vnode.h> 112 113 #include <vm/vm.h> 114 #include <vm/vm_param.h> 115 #include <vm/vm_kern.h> 116 #include <vm/vm_object.h> 117 #include <vm/vm_page.h> 118 #include <vm/vm_pageout.h> 119 #include <vm/vm_pager.h> 120 #include <vm/vm_extern.h> 121 122 /* 123 * Associated with page of user-allocatable memory is a 124 * page structure. 125 */ 126 127 static struct vm_page **vm_page_buckets; /* Array of buckets */ 128 static int vm_page_bucket_count; /* How big is array? */ 129 static int vm_page_hash_mask; /* Mask for hash function */ 130 static volatile int vm_page_bucket_generation; 131 static struct mtx vm_buckets_mtx[BUCKET_HASH_SIZE]; 132 133 vm_page_t vm_page_array = 0; 134 int vm_page_array_size = 0; 135 long first_page = 0; 136 int vm_page_zero_count = 0; 137 138 /* 139 * vm_set_page_size: 140 * 141 * Sets the page size, perhaps based upon the memory 142 * size. Must be called before any use of page-size 143 * dependent functions. 144 */ 145 void 146 vm_set_page_size(void) 147 { 148 if (cnt.v_page_size == 0) 149 cnt.v_page_size = PAGE_SIZE; 150 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 151 panic("vm_set_page_size: page size not a power of two"); 152 } 153 154 /* 155 * vm_page_startup: 156 * 157 * Initializes the resident memory module. 158 * 159 * Allocates memory for the page cells, and 160 * for the object/offset-to-page hash table headers. 161 * Each page cell is initialized and placed on the free list. 162 */ 163 164 vm_offset_t 165 vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr) 166 { 167 vm_offset_t mapped; 168 struct vm_page **bucket; 169 vm_size_t npages, page_range; 170 vm_offset_t new_end; 171 int i; 172 vm_offset_t pa; 173 int nblocks; 174 vm_offset_t last_pa; 175 176 /* the biggest memory array is the second group of pages */ 177 vm_offset_t end; 178 vm_offset_t biggestone, biggestsize; 179 180 vm_offset_t total; 181 182 total = 0; 183 biggestsize = 0; 184 biggestone = 0; 185 nblocks = 0; 186 vaddr = round_page(vaddr); 187 188 for (i = 0; phys_avail[i + 1]; i += 2) { 189 phys_avail[i] = round_page(phys_avail[i]); 190 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 191 } 192 193 for (i = 0; phys_avail[i + 1]; i += 2) { 194 int size = phys_avail[i + 1] - phys_avail[i]; 195 196 if (size > biggestsize) { 197 biggestone = i; 198 biggestsize = size; 199 } 200 ++nblocks; 201 total += size; 202 } 203 204 end = phys_avail[biggestone+1]; 205 206 /* 207 * Initialize the queue headers for the free queue, the active queue 208 * and the inactive queue. 209 */ 210 211 vm_pageq_init(); 212 213 /* 214 * Allocate (and initialize) the hash table buckets. 215 * 216 * The number of buckets MUST BE a power of 2, and the actual value is 217 * the next power of 2 greater than the number of physical pages in 218 * the system. 219 * 220 * We make the hash table approximately 2x the number of pages to 221 * reduce the chain length. This is about the same size using the 222 * singly-linked list as the 1x hash table we were using before 223 * using TAILQ but the chain length will be smaller. 224 * 225 * Note: This computation can be tweaked if desired. 226 */ 227 if (vm_page_bucket_count == 0) { 228 vm_page_bucket_count = 1; 229 while (vm_page_bucket_count < atop(total)) 230 vm_page_bucket_count <<= 1; 231 } 232 vm_page_bucket_count <<= 1; 233 vm_page_hash_mask = vm_page_bucket_count - 1; 234 235 /* 236 * Validate these addresses. 237 */ 238 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *); 239 new_end = trunc_page(new_end); 240 mapped = pmap_map(&vaddr, new_end, end, 241 VM_PROT_READ | VM_PROT_WRITE); 242 bzero((caddr_t) mapped, end - new_end); 243 244 vm_page_buckets = (struct vm_page **)mapped; 245 bucket = vm_page_buckets; 246 for (i = 0; i < vm_page_bucket_count; i++) { 247 *bucket = NULL; 248 bucket++; 249 } 250 for (i = 0; i < BUCKET_HASH_SIZE; ++i) 251 mtx_init(&vm_buckets_mtx[i], "vm buckets hash mutexes", MTX_DEF); 252 253 /* 254 * Compute the number of pages of memory that will be available for 255 * use (taking into account the overhead of a page structure per 256 * page). 257 */ 258 259 first_page = phys_avail[0] / PAGE_SIZE; 260 261 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 262 npages = (total - (page_range * sizeof(struct vm_page)) - 263 (end - new_end)) / PAGE_SIZE; 264 265 end = new_end; 266 267 /* 268 * Initialize the mem entry structures now, and put them in the free 269 * queue. 270 */ 271 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 272 mapped = pmap_map(&vaddr, new_end, end, 273 VM_PROT_READ | VM_PROT_WRITE); 274 vm_page_array = (vm_page_t) mapped; 275 276 /* 277 * Clear all of the page structures 278 */ 279 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 280 vm_page_array_size = page_range; 281 282 /* 283 * Construct the free queue(s) in descending order (by physical 284 * address) so that the first 16MB of physical memory is allocated 285 * last rather than first. On large-memory machines, this avoids 286 * the exhaustion of low physical memory before isa_dmainit has run. 287 */ 288 cnt.v_page_count = 0; 289 cnt.v_free_count = 0; 290 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 291 pa = phys_avail[i]; 292 if (i == biggestone) 293 last_pa = new_end; 294 else 295 last_pa = phys_avail[i + 1]; 296 while (pa < last_pa && npages-- > 0) { 297 vm_pageq_add_new_page(pa); 298 pa += PAGE_SIZE; 299 } 300 } 301 return (vaddr); 302 } 303 304 /* 305 * vm_page_hash: 306 * 307 * Distributes the object/offset key pair among hash buckets. 308 * 309 * NOTE: This macro depends on vm_page_bucket_count being a power of 2. 310 * This routine may not block. 311 * 312 * We try to randomize the hash based on the object to spread the pages 313 * out in the hash table without it costing us too much. 314 */ 315 static __inline int 316 vm_page_hash(vm_object_t object, vm_pindex_t pindex) 317 { 318 int i = ((uintptr_t)object + pindex) ^ object->hash_rand; 319 320 return(i & vm_page_hash_mask); 321 } 322 323 void 324 vm_page_flag_set(vm_page_t m, unsigned short bits) 325 { 326 GIANT_REQUIRED; 327 m->flags |= bits; 328 } 329 330 void 331 vm_page_flag_clear(vm_page_t m, unsigned short bits) 332 { 333 GIANT_REQUIRED; 334 m->flags &= ~bits; 335 } 336 337 void 338 vm_page_busy(vm_page_t m) 339 { 340 KASSERT((m->flags & PG_BUSY) == 0, 341 ("vm_page_busy: page already busy!!!")); 342 vm_page_flag_set(m, PG_BUSY); 343 } 344 345 /* 346 * vm_page_flash: 347 * 348 * wakeup anyone waiting for the page. 349 */ 350 351 void 352 vm_page_flash(vm_page_t m) 353 { 354 if (m->flags & PG_WANTED) { 355 vm_page_flag_clear(m, PG_WANTED); 356 wakeup(m); 357 } 358 } 359 360 /* 361 * vm_page_wakeup: 362 * 363 * clear the PG_BUSY flag and wakeup anyone waiting for the 364 * page. 365 * 366 */ 367 368 void 369 vm_page_wakeup(vm_page_t m) 370 { 371 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 372 vm_page_flag_clear(m, PG_BUSY); 373 vm_page_flash(m); 374 } 375 376 /* 377 * 378 * 379 */ 380 381 void 382 vm_page_io_start(vm_page_t m) 383 { 384 GIANT_REQUIRED; 385 m->busy++; 386 } 387 388 void 389 vm_page_io_finish(vm_page_t m) 390 { 391 GIANT_REQUIRED; 392 m->busy--; 393 if (m->busy == 0) 394 vm_page_flash(m); 395 } 396 397 /* 398 * Keep page from being freed by the page daemon 399 * much of the same effect as wiring, except much lower 400 * overhead and should be used only for *very* temporary 401 * holding ("wiring"). 402 */ 403 void 404 vm_page_hold(vm_page_t mem) 405 { 406 GIANT_REQUIRED; 407 mem->hold_count++; 408 } 409 410 void 411 vm_page_unhold(vm_page_t mem) 412 { 413 GIANT_REQUIRED; 414 --mem->hold_count; 415 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 416 } 417 418 /* 419 * vm_page_protect: 420 * 421 * Reduce the protection of a page. This routine never raises the 422 * protection and therefore can be safely called if the page is already 423 * at VM_PROT_NONE (it will be a NOP effectively ). 424 */ 425 426 void 427 vm_page_protect(vm_page_t mem, int prot) 428 { 429 if (prot == VM_PROT_NONE) { 430 if (mem->flags & (PG_WRITEABLE|PG_MAPPED)) { 431 pmap_page_protect(mem, VM_PROT_NONE); 432 vm_page_flag_clear(mem, PG_WRITEABLE|PG_MAPPED); 433 } 434 } else if ((prot == VM_PROT_READ) && (mem->flags & PG_WRITEABLE)) { 435 pmap_page_protect(mem, VM_PROT_READ); 436 vm_page_flag_clear(mem, PG_WRITEABLE); 437 } 438 } 439 /* 440 * vm_page_zero_fill: 441 * 442 * Zero-fill the specified page. 443 * Written as a standard pagein routine, to 444 * be used by the zero-fill object. 445 */ 446 boolean_t 447 vm_page_zero_fill(vm_page_t m) 448 { 449 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 450 return (TRUE); 451 } 452 453 /* 454 * vm_page_copy: 455 * 456 * Copy one page to another 457 */ 458 void 459 vm_page_copy(vm_page_t src_m, vm_page_t dest_m) 460 { 461 pmap_copy_page(VM_PAGE_TO_PHYS(src_m), VM_PAGE_TO_PHYS(dest_m)); 462 dest_m->valid = VM_PAGE_BITS_ALL; 463 } 464 465 /* 466 * vm_page_free: 467 * 468 * Free a page 469 * 470 * The clearing of PG_ZERO is a temporary safety until the code can be 471 * reviewed to determine that PG_ZERO is being properly cleared on 472 * write faults or maps. PG_ZERO was previously cleared in 473 * vm_page_alloc(). 474 */ 475 void 476 vm_page_free(vm_page_t m) 477 { 478 vm_page_flag_clear(m, PG_ZERO); 479 vm_page_free_toq(m); 480 } 481 482 /* 483 * vm_page_free_zero: 484 * 485 * Free a page to the zerod-pages queue 486 */ 487 void 488 vm_page_free_zero(vm_page_t m) 489 { 490 vm_page_flag_set(m, PG_ZERO); 491 vm_page_free_toq(m); 492 } 493 494 /* 495 * vm_page_sleep_busy: 496 * 497 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE) 498 * m->busy is zero. Returns TRUE if it had to sleep ( including if 499 * it almost had to sleep and made temporary spl*() mods), FALSE 500 * otherwise. 501 * 502 * This routine assumes that interrupts can only remove the busy 503 * status from a page, not set the busy status or change it from 504 * PG_BUSY to m->busy or vise versa (which would create a timing 505 * window). 506 */ 507 508 int 509 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 510 { 511 GIANT_REQUIRED; 512 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 513 int s = splvm(); 514 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 515 /* 516 * Page is busy. Wait and retry. 517 */ 518 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 519 tsleep(m, PVM, msg, 0); 520 } 521 splx(s); 522 return(TRUE); 523 /* not reached */ 524 } 525 return(FALSE); 526 } 527 /* 528 * vm_page_dirty: 529 * 530 * make page all dirty 531 */ 532 533 void 534 vm_page_dirty(vm_page_t m) 535 { 536 KASSERT(m->queue - m->pc != PQ_CACHE, 537 ("vm_page_dirty: page in cache!")); 538 m->dirty = VM_PAGE_BITS_ALL; 539 } 540 541 /* 542 * vm_page_undirty: 543 * 544 * Set page to not be dirty. Note: does not clear pmap modify bits 545 */ 546 547 void 548 vm_page_undirty(vm_page_t m) 549 { 550 m->dirty = 0; 551 } 552 553 /* 554 * vm_page_insert: [ internal use only ] 555 * 556 * Inserts the given mem entry into the object and object list. 557 * 558 * The pagetables are not updated but will presumably fault the page 559 * in if necessary, or if a kernel page the caller will at some point 560 * enter the page into the kernel's pmap. We are not allowed to block 561 * here so we *can't* do this anyway. 562 * 563 * The object and page must be locked, and must be splhigh. 564 * This routine may not block. 565 */ 566 567 void 568 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 569 { 570 struct vm_page **bucket; 571 572 GIANT_REQUIRED; 573 574 if (m->object != NULL) 575 panic("vm_page_insert: already inserted"); 576 577 /* 578 * Record the object/offset pair in this page 579 */ 580 581 m->object = object; 582 m->pindex = pindex; 583 584 /* 585 * Insert it into the object_object/offset hash table 586 */ 587 588 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 589 m->hnext = *bucket; 590 *bucket = m; 591 vm_page_bucket_generation++; 592 593 /* 594 * Now link into the object's list of backed pages. 595 */ 596 597 TAILQ_INSERT_TAIL(&object->memq, m, listq); 598 object->generation++; 599 600 /* 601 * show that the object has one more resident page. 602 */ 603 604 object->resident_page_count++; 605 606 /* 607 * Since we are inserting a new and possibly dirty page, 608 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 609 */ 610 if (m->flags & PG_WRITEABLE) 611 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 612 } 613 614 /* 615 * vm_page_remove: 616 * NOTE: used by device pager as well -wfj 617 * 618 * Removes the given mem entry from the object/offset-page 619 * table and the object page list, but do not invalidate/terminate 620 * the backing store. 621 * 622 * The object and page must be locked, and at splhigh. 623 * The underlying pmap entry (if any) is NOT removed here. 624 * This routine may not block. 625 */ 626 627 void 628 vm_page_remove(vm_page_t m) 629 { 630 vm_object_t object; 631 632 GIANT_REQUIRED; 633 634 if (m->object == NULL) 635 return; 636 637 if ((m->flags & PG_BUSY) == 0) { 638 panic("vm_page_remove: page not busy"); 639 } 640 641 /* 642 * Basically destroy the page. 643 */ 644 645 vm_page_wakeup(m); 646 647 object = m->object; 648 649 /* 650 * Remove from the object_object/offset hash table. The object 651 * must be on the hash queue, we will panic if it isn't 652 * 653 * Note: we must NULL-out m->hnext to prevent loops in detached 654 * buffers with vm_page_lookup(). 655 */ 656 657 { 658 struct vm_page **bucket; 659 660 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)]; 661 while (*bucket != m) { 662 if (*bucket == NULL) 663 panic("vm_page_remove(): page not found in hash"); 664 bucket = &(*bucket)->hnext; 665 } 666 *bucket = m->hnext; 667 m->hnext = NULL; 668 vm_page_bucket_generation++; 669 } 670 671 /* 672 * Now remove from the object's list of backed pages. 673 */ 674 675 TAILQ_REMOVE(&object->memq, m, listq); 676 677 /* 678 * And show that the object has one fewer resident page. 679 */ 680 681 object->resident_page_count--; 682 object->generation++; 683 684 m->object = NULL; 685 } 686 687 /* 688 * vm_page_lookup: 689 * 690 * Returns the page associated with the object/offset 691 * pair specified; if none is found, NULL is returned. 692 * 693 * NOTE: the code below does not lock. It will operate properly if 694 * an interrupt makes a change, but the generation algorithm will not 695 * operate properly in an SMP environment where both cpu's are able to run 696 * kernel code simultaneously. 697 * 698 * The object must be locked. No side effects. 699 * This routine may not block. 700 * This is a critical path routine 701 */ 702 703 vm_page_t 704 vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 705 { 706 vm_page_t m; 707 struct vm_page **bucket; 708 int generation; 709 710 /* 711 * Search the hash table for this object/offset pair 712 */ 713 714 retry: 715 generation = vm_page_bucket_generation; 716 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 717 for (m = *bucket; m != NULL; m = m->hnext) { 718 if ((m->object == object) && (m->pindex == pindex)) { 719 if (vm_page_bucket_generation != generation) 720 goto retry; 721 return (m); 722 } 723 } 724 if (vm_page_bucket_generation != generation) 725 goto retry; 726 return (NULL); 727 } 728 729 /* 730 * vm_page_rename: 731 * 732 * Move the given memory entry from its 733 * current object to the specified target object/offset. 734 * 735 * The object must be locked. 736 * This routine may not block. 737 * 738 * Note: this routine will raise itself to splvm(), the caller need not. 739 * 740 * Note: swap associated with the page must be invalidated by the move. We 741 * have to do this for several reasons: (1) we aren't freeing the 742 * page, (2) we are dirtying the page, (3) the VM system is probably 743 * moving the page from object A to B, and will then later move 744 * the backing store from A to B and we can't have a conflict. 745 * 746 * Note: we *always* dirty the page. It is necessary both for the 747 * fact that we moved it, and because we may be invalidating 748 * swap. If the page is on the cache, we have to deactivate it 749 * or vm_page_dirty() will panic. Dirty pages are not allowed 750 * on the cache. 751 */ 752 753 void 754 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 755 { 756 int s; 757 758 s = splvm(); 759 vm_page_remove(m); 760 vm_page_insert(m, new_object, new_pindex); 761 if (m->queue - m->pc == PQ_CACHE) 762 vm_page_deactivate(m); 763 vm_page_dirty(m); 764 splx(s); 765 } 766 767 /* 768 * vm_page_select_cache: 769 * 770 * Find a page on the cache queue with color optimization. As pages 771 * might be found, but not applicable, they are deactivated. This 772 * keeps us from using potentially busy cached pages. 773 * 774 * This routine must be called at splvm(). 775 * This routine may not block. 776 */ 777 static vm_page_t 778 vm_page_select_cache(vm_object_t object, vm_pindex_t pindex) 779 { 780 vm_page_t m; 781 782 GIANT_REQUIRED; 783 while (TRUE) { 784 m = vm_pageq_find( 785 PQ_CACHE, 786 (pindex + object->pg_color) & PQ_L2_MASK, 787 FALSE 788 ); 789 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 790 m->hold_count || m->wire_count)) { 791 vm_page_deactivate(m); 792 continue; 793 } 794 return m; 795 } 796 } 797 798 /* 799 * vm_page_select_free: 800 * 801 * Find a free or zero page, with specified preference. 802 * 803 * This routine must be called at splvm(). 804 * This routine may not block. 805 */ 806 807 static __inline vm_page_t 808 vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 809 { 810 vm_page_t m; 811 812 m = vm_pageq_find( 813 PQ_FREE, 814 (pindex + object->pg_color) & PQ_L2_MASK, 815 prefer_zero 816 ); 817 return(m); 818 } 819 820 /* 821 * vm_page_alloc: 822 * 823 * Allocate and return a memory cell associated 824 * with this VM object/offset pair. 825 * 826 * page_req classes: 827 * VM_ALLOC_NORMAL normal process request 828 * VM_ALLOC_SYSTEM system *really* needs a page 829 * VM_ALLOC_INTERRUPT interrupt time request 830 * VM_ALLOC_ZERO zero page 831 * 832 * This routine may not block. 833 * 834 * Additional special handling is required when called from an 835 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 836 * the page cache in this case. 837 */ 838 839 vm_page_t 840 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 841 { 842 vm_page_t m = NULL; 843 int s; 844 845 GIANT_REQUIRED; 846 847 KASSERT(!vm_page_lookup(object, pindex), 848 ("vm_page_alloc: page already allocated")); 849 850 /* 851 * The pager is allowed to eat deeper into the free page list. 852 */ 853 854 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 855 page_req = VM_ALLOC_SYSTEM; 856 }; 857 858 s = splvm(); 859 860 loop: 861 if (cnt.v_free_count > cnt.v_free_reserved) { 862 /* 863 * Allocate from the free queue if there are plenty of pages 864 * in it. 865 */ 866 if (page_req == VM_ALLOC_ZERO) 867 m = vm_page_select_free(object, pindex, TRUE); 868 else 869 m = vm_page_select_free(object, pindex, FALSE); 870 } else if ( 871 (page_req == VM_ALLOC_SYSTEM && 872 cnt.v_cache_count == 0 && 873 cnt.v_free_count > cnt.v_interrupt_free_min) || 874 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0) 875 ) { 876 /* 877 * Interrupt or system, dig deeper into the free list. 878 */ 879 m = vm_page_select_free(object, pindex, FALSE); 880 } else if (page_req != VM_ALLOC_INTERRUPT) { 881 /* 882 * Allocatable from cache (non-interrupt only). On success, 883 * we must free the page and try again, thus ensuring that 884 * cnt.v_*_free_min counters are replenished. 885 */ 886 m = vm_page_select_cache(object, pindex); 887 if (m == NULL) { 888 splx(s); 889 #if defined(DIAGNOSTIC) 890 if (cnt.v_cache_count > 0) 891 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 892 #endif 893 vm_pageout_deficit++; 894 pagedaemon_wakeup(); 895 return (NULL); 896 } 897 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 898 vm_page_busy(m); 899 vm_page_protect(m, VM_PROT_NONE); 900 vm_page_free(m); 901 goto loop; 902 } else { 903 /* 904 * Not allocatable from cache from interrupt, give up. 905 */ 906 splx(s); 907 vm_pageout_deficit++; 908 pagedaemon_wakeup(); 909 return (NULL); 910 } 911 912 /* 913 * At this point we had better have found a good page. 914 */ 915 916 KASSERT( 917 m != NULL, 918 ("vm_page_alloc(): missing page on free queue\n") 919 ); 920 921 /* 922 * Remove from free queue 923 */ 924 925 vm_pageq_remove_nowakeup(m); 926 927 /* 928 * Initialize structure. Only the PG_ZERO flag is inherited. 929 */ 930 931 if (m->flags & PG_ZERO) { 932 vm_page_zero_count--; 933 m->flags = PG_ZERO | PG_BUSY; 934 } else { 935 m->flags = PG_BUSY; 936 } 937 m->wire_count = 0; 938 m->hold_count = 0; 939 m->act_count = 0; 940 m->busy = 0; 941 m->valid = 0; 942 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 943 944 /* 945 * vm_page_insert() is safe prior to the splx(). Note also that 946 * inserting a page here does not insert it into the pmap (which 947 * could cause us to block allocating memory). We cannot block 948 * anywhere. 949 */ 950 951 vm_page_insert(m, object, pindex); 952 953 /* 954 * Don't wakeup too often - wakeup the pageout daemon when 955 * we would be nearly out of memory. 956 */ 957 if (vm_paging_needed()) 958 pagedaemon_wakeup(); 959 960 splx(s); 961 962 return (m); 963 } 964 965 /* 966 * vm_wait: (also see VM_WAIT macro) 967 * 968 * Block until free pages are available for allocation 969 */ 970 971 void 972 vm_wait(void) 973 { 974 int s; 975 976 s = splvm(); 977 if (curproc == pageproc) { 978 vm_pageout_pages_needed = 1; 979 tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0); 980 } else { 981 if (!vm_pages_needed) { 982 vm_pages_needed = 1; 983 wakeup(&vm_pages_needed); 984 } 985 tsleep(&cnt.v_free_count, PVM, "vmwait", 0); 986 } 987 splx(s); 988 } 989 990 /* 991 * vm_await: (also see VM_AWAIT macro) 992 * 993 * asleep on an event that will signal when free pages are available 994 * for allocation. 995 */ 996 997 void 998 vm_await(void) 999 { 1000 int s; 1001 1002 s = splvm(); 1003 if (curproc == pageproc) { 1004 vm_pageout_pages_needed = 1; 1005 asleep(&vm_pageout_pages_needed, PSWP, "vmwait", 0); 1006 } else { 1007 if (!vm_pages_needed) { 1008 vm_pages_needed++; 1009 wakeup(&vm_pages_needed); 1010 } 1011 asleep(&cnt.v_free_count, PVM, "vmwait", 0); 1012 } 1013 splx(s); 1014 } 1015 1016 /* 1017 * vm_page_activate: 1018 * 1019 * Put the specified page on the active list (if appropriate). 1020 * Ensure that act_count is at least ACT_INIT but do not otherwise 1021 * mess with it. 1022 * 1023 * The page queues must be locked. 1024 * This routine may not block. 1025 */ 1026 void 1027 vm_page_activate(vm_page_t m) 1028 { 1029 int s; 1030 1031 GIANT_REQUIRED; 1032 s = splvm(); 1033 1034 if (m->queue != PQ_ACTIVE) { 1035 if ((m->queue - m->pc) == PQ_CACHE) 1036 cnt.v_reactivated++; 1037 1038 vm_pageq_remove(m); 1039 1040 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1041 m->queue = PQ_ACTIVE; 1042 vm_page_queues[PQ_ACTIVE].lcnt++; 1043 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1044 if (m->act_count < ACT_INIT) 1045 m->act_count = ACT_INIT; 1046 cnt.v_active_count++; 1047 } 1048 } else { 1049 if (m->act_count < ACT_INIT) 1050 m->act_count = ACT_INIT; 1051 } 1052 1053 splx(s); 1054 } 1055 1056 /* 1057 * vm_page_free_wakeup: 1058 * 1059 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1060 * routine is called when a page has been added to the cache or free 1061 * queues. 1062 * 1063 * This routine may not block. 1064 * This routine must be called at splvm() 1065 */ 1066 static __inline void 1067 vm_page_free_wakeup(void) 1068 { 1069 /* 1070 * if pageout daemon needs pages, then tell it that there are 1071 * some free. 1072 */ 1073 if (vm_pageout_pages_needed && 1074 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1075 wakeup(&vm_pageout_pages_needed); 1076 vm_pageout_pages_needed = 0; 1077 } 1078 /* 1079 * wakeup processes that are waiting on memory if we hit a 1080 * high water mark. And wakeup scheduler process if we have 1081 * lots of memory. this process will swapin processes. 1082 */ 1083 if (vm_pages_needed && !vm_page_count_min()) { 1084 vm_pages_needed = 0; 1085 wakeup(&cnt.v_free_count); 1086 } 1087 } 1088 1089 /* 1090 * vm_page_free_toq: 1091 * 1092 * Returns the given page to the PQ_FREE list, 1093 * disassociating it with any VM object. 1094 * 1095 * Object and page must be locked prior to entry. 1096 * This routine may not block. 1097 */ 1098 1099 void 1100 vm_page_free_toq(vm_page_t m) 1101 { 1102 int s; 1103 struct vpgqueues *pq; 1104 vm_object_t object = m->object; 1105 1106 GIANT_REQUIRED; 1107 s = splvm(); 1108 cnt.v_tfree++; 1109 1110 if (m->busy || ((m->queue - m->pc) == PQ_FREE) || 1111 (m->hold_count != 0)) { 1112 printf( 1113 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1114 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1115 m->hold_count); 1116 if ((m->queue - m->pc) == PQ_FREE) 1117 panic("vm_page_free: freeing free page"); 1118 else 1119 panic("vm_page_free: freeing busy page"); 1120 } 1121 1122 /* 1123 * unqueue, then remove page. Note that we cannot destroy 1124 * the page here because we do not want to call the pager's 1125 * callback routine until after we've put the page on the 1126 * appropriate free queue. 1127 */ 1128 1129 vm_pageq_remove_nowakeup(m); 1130 vm_page_remove(m); 1131 1132 /* 1133 * If fictitious remove object association and 1134 * return, otherwise delay object association removal. 1135 */ 1136 1137 if ((m->flags & PG_FICTITIOUS) != 0) { 1138 splx(s); 1139 return; 1140 } 1141 1142 m->valid = 0; 1143 vm_page_undirty(m); 1144 1145 if (m->wire_count != 0) { 1146 if (m->wire_count > 1) { 1147 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1148 m->wire_count, (long)m->pindex); 1149 } 1150 panic("vm_page_free: freeing wired page\n"); 1151 } 1152 1153 /* 1154 * If we've exhausted the object's resident pages we want to free 1155 * it up. 1156 */ 1157 1158 if (object && 1159 (object->type == OBJT_VNODE) && 1160 ((object->flags & OBJ_DEAD) == 0) 1161 ) { 1162 struct vnode *vp = (struct vnode *)object->handle; 1163 1164 if (vp && VSHOULDFREE(vp)) 1165 vfree(vp); 1166 } 1167 1168 /* 1169 * Clear the UNMANAGED flag when freeing an unmanaged page. 1170 */ 1171 1172 if (m->flags & PG_UNMANAGED) { 1173 m->flags &= ~PG_UNMANAGED; 1174 } else { 1175 #ifdef __alpha__ 1176 pmap_page_is_free(m); 1177 #endif 1178 } 1179 1180 m->queue = PQ_FREE + m->pc; 1181 pq = &vm_page_queues[m->queue]; 1182 pq->lcnt++; 1183 ++(*pq->cnt); 1184 1185 /* 1186 * Put zero'd pages on the end ( where we look for zero'd pages 1187 * first ) and non-zerod pages at the head. 1188 */ 1189 1190 if (m->flags & PG_ZERO) { 1191 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1192 ++vm_page_zero_count; 1193 } else { 1194 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1195 } 1196 1197 vm_page_free_wakeup(); 1198 1199 splx(s); 1200 } 1201 1202 /* 1203 * vm_page_unmanage: 1204 * 1205 * Prevent PV management from being done on the page. The page is 1206 * removed from the paging queues as if it were wired, and as a 1207 * consequence of no longer being managed the pageout daemon will not 1208 * touch it (since there is no way to locate the pte mappings for the 1209 * page). madvise() calls that mess with the pmap will also no longer 1210 * operate on the page. 1211 * 1212 * Beyond that the page is still reasonably 'normal'. Freeing the page 1213 * will clear the flag. 1214 * 1215 * This routine is used by OBJT_PHYS objects - objects using unswappable 1216 * physical memory as backing store rather then swap-backed memory and 1217 * will eventually be extended to support 4MB unmanaged physical 1218 * mappings. 1219 */ 1220 1221 void 1222 vm_page_unmanage(vm_page_t m) 1223 { 1224 int s; 1225 1226 s = splvm(); 1227 if ((m->flags & PG_UNMANAGED) == 0) { 1228 if (m->wire_count == 0) 1229 vm_pageq_remove(m); 1230 } 1231 vm_page_flag_set(m, PG_UNMANAGED); 1232 splx(s); 1233 } 1234 1235 /* 1236 * vm_page_wire: 1237 * 1238 * Mark this page as wired down by yet 1239 * another map, removing it from paging queues 1240 * as necessary. 1241 * 1242 * The page queues must be locked. 1243 * This routine may not block. 1244 */ 1245 void 1246 vm_page_wire(vm_page_t m) 1247 { 1248 int s; 1249 1250 /* 1251 * Only bump the wire statistics if the page is not already wired, 1252 * and only unqueue the page if it is on some queue (if it is unmanaged 1253 * it is already off the queues). 1254 */ 1255 s = splvm(); 1256 if (m->wire_count == 0) { 1257 if ((m->flags & PG_UNMANAGED) == 0) 1258 vm_pageq_remove(m); 1259 cnt.v_wire_count++; 1260 } 1261 m->wire_count++; 1262 splx(s); 1263 vm_page_flag_set(m, PG_MAPPED); 1264 } 1265 1266 /* 1267 * vm_page_unwire: 1268 * 1269 * Release one wiring of this page, potentially 1270 * enabling it to be paged again. 1271 * 1272 * Many pages placed on the inactive queue should actually go 1273 * into the cache, but it is difficult to figure out which. What 1274 * we do instead, if the inactive target is well met, is to put 1275 * clean pages at the head of the inactive queue instead of the tail. 1276 * This will cause them to be moved to the cache more quickly and 1277 * if not actively re-referenced, freed more quickly. If we just 1278 * stick these pages at the end of the inactive queue, heavy filesystem 1279 * meta-data accesses can cause an unnecessary paging load on memory bound 1280 * processes. This optimization causes one-time-use metadata to be 1281 * reused more quickly. 1282 * 1283 * BUT, if we are in a low-memory situation we have no choice but to 1284 * put clean pages on the cache queue. 1285 * 1286 * A number of routines use vm_page_unwire() to guarantee that the page 1287 * will go into either the inactive or active queues, and will NEVER 1288 * be placed in the cache - for example, just after dirtying a page. 1289 * dirty pages in the cache are not allowed. 1290 * 1291 * The page queues must be locked. 1292 * This routine may not block. 1293 */ 1294 void 1295 vm_page_unwire(vm_page_t m, int activate) 1296 { 1297 int s; 1298 1299 s = splvm(); 1300 1301 if (m->wire_count > 0) { 1302 m->wire_count--; 1303 if (m->wire_count == 0) { 1304 cnt.v_wire_count--; 1305 if (m->flags & PG_UNMANAGED) { 1306 ; 1307 } else if (activate) { 1308 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1309 m->queue = PQ_ACTIVE; 1310 vm_page_queues[PQ_ACTIVE].lcnt++; 1311 cnt.v_active_count++; 1312 } else { 1313 vm_page_flag_clear(m, PG_WINATCFLS); 1314 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1315 m->queue = PQ_INACTIVE; 1316 vm_page_queues[PQ_INACTIVE].lcnt++; 1317 cnt.v_inactive_count++; 1318 } 1319 } 1320 } else { 1321 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1322 } 1323 splx(s); 1324 } 1325 1326 1327 /* 1328 * Move the specified page to the inactive queue. If the page has 1329 * any associated swap, the swap is deallocated. 1330 * 1331 * Normally athead is 0 resulting in LRU operation. athead is set 1332 * to 1 if we want this page to be 'as if it were placed in the cache', 1333 * except without unmapping it from the process address space. 1334 * 1335 * This routine may not block. 1336 */ 1337 static __inline void 1338 _vm_page_deactivate(vm_page_t m, int athead) 1339 { 1340 int s; 1341 1342 GIANT_REQUIRED; 1343 /* 1344 * Ignore if already inactive. 1345 */ 1346 if (m->queue == PQ_INACTIVE) 1347 return; 1348 1349 s = splvm(); 1350 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1351 if ((m->queue - m->pc) == PQ_CACHE) 1352 cnt.v_reactivated++; 1353 vm_page_flag_clear(m, PG_WINATCFLS); 1354 vm_pageq_remove(m); 1355 if (athead) 1356 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1357 else 1358 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1359 m->queue = PQ_INACTIVE; 1360 vm_page_queues[PQ_INACTIVE].lcnt++; 1361 cnt.v_inactive_count++; 1362 } 1363 splx(s); 1364 } 1365 1366 void 1367 vm_page_deactivate(vm_page_t m) 1368 { 1369 _vm_page_deactivate(m, 0); 1370 } 1371 1372 /* 1373 * vm_page_try_to_cache: 1374 * 1375 * Returns 0 on failure, 1 on success 1376 */ 1377 int 1378 vm_page_try_to_cache(vm_page_t m) 1379 { 1380 GIANT_REQUIRED; 1381 1382 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1383 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1384 return(0); 1385 } 1386 vm_page_test_dirty(m); 1387 if (m->dirty) 1388 return(0); 1389 vm_page_cache(m); 1390 return(1); 1391 } 1392 1393 /* 1394 * vm_page_try_to_free() 1395 * 1396 * Attempt to free the page. If we cannot free it, we do nothing. 1397 * 1 is returned on success, 0 on failure. 1398 */ 1399 int 1400 vm_page_try_to_free(vm_page_t m) 1401 { 1402 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1403 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1404 return(0); 1405 } 1406 vm_page_test_dirty(m); 1407 if (m->dirty) 1408 return(0); 1409 vm_page_busy(m); 1410 vm_page_protect(m, VM_PROT_NONE); 1411 vm_page_free(m); 1412 return(1); 1413 } 1414 1415 /* 1416 * vm_page_cache 1417 * 1418 * Put the specified page onto the page cache queue (if appropriate). 1419 * 1420 * This routine may not block. 1421 */ 1422 void 1423 vm_page_cache(vm_page_t m) 1424 { 1425 int s; 1426 1427 GIANT_REQUIRED; 1428 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) { 1429 printf("vm_page_cache: attempting to cache busy page\n"); 1430 return; 1431 } 1432 if ((m->queue - m->pc) == PQ_CACHE) 1433 return; 1434 1435 /* 1436 * Remove all pmaps and indicate that the page is not 1437 * writeable or mapped. 1438 */ 1439 1440 vm_page_protect(m, VM_PROT_NONE); 1441 if (m->dirty != 0) { 1442 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1443 (long)m->pindex); 1444 } 1445 s = splvm(); 1446 vm_pageq_remove_nowakeup(m); 1447 m->queue = PQ_CACHE + m->pc; 1448 vm_page_queues[m->queue].lcnt++; 1449 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 1450 cnt.v_cache_count++; 1451 vm_page_free_wakeup(); 1452 splx(s); 1453 } 1454 1455 /* 1456 * vm_page_dontneed 1457 * 1458 * Cache, deactivate, or do nothing as appropriate. This routine 1459 * is typically used by madvise() MADV_DONTNEED. 1460 * 1461 * Generally speaking we want to move the page into the cache so 1462 * it gets reused quickly. However, this can result in a silly syndrome 1463 * due to the page recycling too quickly. Small objects will not be 1464 * fully cached. On the otherhand, if we move the page to the inactive 1465 * queue we wind up with a problem whereby very large objects 1466 * unnecessarily blow away our inactive and cache queues. 1467 * 1468 * The solution is to move the pages based on a fixed weighting. We 1469 * either leave them alone, deactivate them, or move them to the cache, 1470 * where moving them to the cache has the highest weighting. 1471 * By forcing some pages into other queues we eventually force the 1472 * system to balance the queues, potentially recovering other unrelated 1473 * space from active. The idea is to not force this to happen too 1474 * often. 1475 */ 1476 1477 void 1478 vm_page_dontneed(vm_page_t m) 1479 { 1480 static int dnweight; 1481 int dnw; 1482 int head; 1483 1484 GIANT_REQUIRED; 1485 dnw = ++dnweight; 1486 1487 /* 1488 * occassionally leave the page alone 1489 */ 1490 1491 if ((dnw & 0x01F0) == 0 || 1492 m->queue == PQ_INACTIVE || 1493 m->queue - m->pc == PQ_CACHE 1494 ) { 1495 if (m->act_count >= ACT_INIT) 1496 --m->act_count; 1497 return; 1498 } 1499 1500 if (m->dirty == 0) 1501 vm_page_test_dirty(m); 1502 1503 if (m->dirty || (dnw & 0x0070) == 0) { 1504 /* 1505 * Deactivate the page 3 times out of 32. 1506 */ 1507 head = 0; 1508 } else { 1509 /* 1510 * Cache the page 28 times out of every 32. Note that 1511 * the page is deactivated instead of cached, but placed 1512 * at the head of the queue instead of the tail. 1513 */ 1514 head = 1; 1515 } 1516 _vm_page_deactivate(m, head); 1517 } 1518 1519 /* 1520 * Grab a page, waiting until we are waken up due to the page 1521 * changing state. We keep on waiting, if the page continues 1522 * to be in the object. If the page doesn't exist, allocate it. 1523 * 1524 * This routine may block. 1525 */ 1526 vm_page_t 1527 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1528 { 1529 vm_page_t m; 1530 int s, generation; 1531 1532 GIANT_REQUIRED; 1533 retrylookup: 1534 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1535 if (m->busy || (m->flags & PG_BUSY)) { 1536 generation = object->generation; 1537 1538 s = splvm(); 1539 while ((object->generation == generation) && 1540 (m->busy || (m->flags & PG_BUSY))) { 1541 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1542 tsleep(m, PVM, "pgrbwt", 0); 1543 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1544 splx(s); 1545 return NULL; 1546 } 1547 } 1548 splx(s); 1549 goto retrylookup; 1550 } else { 1551 vm_page_busy(m); 1552 return m; 1553 } 1554 } 1555 1556 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1557 if (m == NULL) { 1558 VM_WAIT; 1559 if ((allocflags & VM_ALLOC_RETRY) == 0) 1560 return NULL; 1561 goto retrylookup; 1562 } 1563 1564 return m; 1565 } 1566 1567 /* 1568 * Mapping function for valid bits or for dirty bits in 1569 * a page. May not block. 1570 * 1571 * Inputs are required to range within a page. 1572 */ 1573 1574 __inline int 1575 vm_page_bits(int base, int size) 1576 { 1577 int first_bit; 1578 int last_bit; 1579 1580 KASSERT( 1581 base + size <= PAGE_SIZE, 1582 ("vm_page_bits: illegal base/size %d/%d", base, size) 1583 ); 1584 1585 if (size == 0) /* handle degenerate case */ 1586 return(0); 1587 1588 first_bit = base >> DEV_BSHIFT; 1589 last_bit = (base + size - 1) >> DEV_BSHIFT; 1590 1591 return ((2 << last_bit) - (1 << first_bit)); 1592 } 1593 1594 /* 1595 * vm_page_set_validclean: 1596 * 1597 * Sets portions of a page valid and clean. The arguments are expected 1598 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1599 * of any partial chunks touched by the range. The invalid portion of 1600 * such chunks will be zero'd. 1601 * 1602 * This routine may not block. 1603 * 1604 * (base + size) must be less then or equal to PAGE_SIZE. 1605 */ 1606 void 1607 vm_page_set_validclean(vm_page_t m, int base, int size) 1608 { 1609 int pagebits; 1610 int frag; 1611 int endoff; 1612 1613 GIANT_REQUIRED; 1614 if (size == 0) /* handle degenerate case */ 1615 return; 1616 1617 /* 1618 * If the base is not DEV_BSIZE aligned and the valid 1619 * bit is clear, we have to zero out a portion of the 1620 * first block. 1621 */ 1622 1623 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1624 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 1625 ) { 1626 pmap_zero_page_area( 1627 VM_PAGE_TO_PHYS(m), 1628 frag, 1629 base - frag 1630 ); 1631 } 1632 1633 /* 1634 * If the ending offset is not DEV_BSIZE aligned and the 1635 * valid bit is clear, we have to zero out a portion of 1636 * the last block. 1637 */ 1638 1639 endoff = base + size; 1640 1641 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1642 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 1643 ) { 1644 pmap_zero_page_area( 1645 VM_PAGE_TO_PHYS(m), 1646 endoff, 1647 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 1648 ); 1649 } 1650 1651 /* 1652 * Set valid, clear dirty bits. If validating the entire 1653 * page we can safely clear the pmap modify bit. We also 1654 * use this opportunity to clear the PG_NOSYNC flag. If a process 1655 * takes a write fault on a MAP_NOSYNC memory area the flag will 1656 * be set again. 1657 */ 1658 1659 pagebits = vm_page_bits(base, size); 1660 m->valid |= pagebits; 1661 m->dirty &= ~pagebits; 1662 if (base == 0 && size == PAGE_SIZE) { 1663 pmap_clear_modify(m); 1664 vm_page_flag_clear(m, PG_NOSYNC); 1665 } 1666 } 1667 1668 #if 0 1669 1670 void 1671 vm_page_set_dirty(vm_page_t m, int base, int size) 1672 { 1673 m->dirty |= vm_page_bits(base, size); 1674 } 1675 1676 #endif 1677 1678 void 1679 vm_page_clear_dirty(vm_page_t m, int base, int size) 1680 { 1681 GIANT_REQUIRED; 1682 m->dirty &= ~vm_page_bits(base, size); 1683 } 1684 1685 /* 1686 * vm_page_set_invalid: 1687 * 1688 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1689 * valid and dirty bits for the effected areas are cleared. 1690 * 1691 * May not block. 1692 */ 1693 void 1694 vm_page_set_invalid(vm_page_t m, int base, int size) 1695 { 1696 int bits; 1697 1698 GIANT_REQUIRED; 1699 bits = vm_page_bits(base, size); 1700 m->valid &= ~bits; 1701 m->dirty &= ~bits; 1702 m->object->generation++; 1703 } 1704 1705 /* 1706 * vm_page_zero_invalid() 1707 * 1708 * The kernel assumes that the invalid portions of a page contain 1709 * garbage, but such pages can be mapped into memory by user code. 1710 * When this occurs, we must zero out the non-valid portions of the 1711 * page so user code sees what it expects. 1712 * 1713 * Pages are most often semi-valid when the end of a file is mapped 1714 * into memory and the file's size is not page aligned. 1715 */ 1716 1717 void 1718 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1719 { 1720 int b; 1721 int i; 1722 1723 /* 1724 * Scan the valid bits looking for invalid sections that 1725 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1726 * valid bit may be set ) have already been zerod by 1727 * vm_page_set_validclean(). 1728 */ 1729 1730 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1731 if (i == (PAGE_SIZE / DEV_BSIZE) || 1732 (m->valid & (1 << i)) 1733 ) { 1734 if (i > b) { 1735 pmap_zero_page_area( 1736 VM_PAGE_TO_PHYS(m), 1737 b << DEV_BSHIFT, 1738 (i - b) << DEV_BSHIFT 1739 ); 1740 } 1741 b = i + 1; 1742 } 1743 } 1744 1745 /* 1746 * setvalid is TRUE when we can safely set the zero'd areas 1747 * as being valid. We can do this if there are no cache consistancy 1748 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1749 */ 1750 1751 if (setvalid) 1752 m->valid = VM_PAGE_BITS_ALL; 1753 } 1754 1755 /* 1756 * vm_page_is_valid: 1757 * 1758 * Is (partial) page valid? Note that the case where size == 0 1759 * will return FALSE in the degenerate case where the page is 1760 * entirely invalid, and TRUE otherwise. 1761 * 1762 * May not block. 1763 */ 1764 1765 int 1766 vm_page_is_valid(vm_page_t m, int base, int size) 1767 { 1768 int bits = vm_page_bits(base, size); 1769 1770 if (m->valid && ((m->valid & bits) == bits)) 1771 return 1; 1772 else 1773 return 0; 1774 } 1775 1776 /* 1777 * update dirty bits from pmap/mmu. May not block. 1778 */ 1779 1780 void 1781 vm_page_test_dirty(vm_page_t m) 1782 { 1783 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1784 vm_page_dirty(m); 1785 } 1786 } 1787 1788 #include "opt_ddb.h" 1789 #ifdef DDB 1790 #include <sys/kernel.h> 1791 1792 #include <ddb/ddb.h> 1793 1794 DB_SHOW_COMMAND(page, vm_page_print_page_info) 1795 { 1796 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1797 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1798 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1799 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1800 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1801 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1802 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1803 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1804 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1805 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1806 } 1807 1808 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1809 { 1810 int i; 1811 db_printf("PQ_FREE:"); 1812 for (i = 0; i < PQ_L2_SIZE; i++) { 1813 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1814 } 1815 db_printf("\n"); 1816 1817 db_printf("PQ_CACHE:"); 1818 for (i = 0; i < PQ_L2_SIZE; i++) { 1819 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1820 } 1821 db_printf("\n"); 1822 1823 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1824 vm_page_queues[PQ_ACTIVE].lcnt, 1825 vm_page_queues[PQ_INACTIVE].lcnt); 1826 } 1827 #endif /* DDB */ 1828